Not exactly the best experimental design here. Instead, canadian cross should be used which would reduce and detect variation between individuals.

(Same devs on multiple occasions, sometimes well slept, sometimes not. Matching tasks for both groups. Learning effect is controlled by swapping group order of task execution. This is fully factorial fully crossed design.)

However, the magnitude is indeed big and the results probably valid but might not quite match up for experienced devs.